GWPV-YOLO: Real-time Photovoltaic Defect Detection Based on UAV Thermal Imaging

Unmanned Aerial Vehicles (UAVs) have become crucial tools for photovoltaic (PV) panel defect detection. However, traditional Convolutional Neural Network (CNN) models face challenges in detecting tiny hotspots with irregular shapes in thermal images due to complex backgrounds, leading to misidentification of abnormal heat sources from non-PV panels as defects. Additionally, small-target defect information is prone to loss during detection, causing false or missed detections, and large model sizes often compromise real-time performance. To address these issues, this paper proposes a novel lightweight detection network, GWPV-YOLO, improved based on YOLO11. The network first replaces the C2PSA attention module in YOLO11 with a Global Grouped Coordinate Attention (GGCA) module to capture multi-dimensional global information, enhancing feature extraction comprehensiveness, reducing misidentification of nonPV heat sources, and improving small-target defect detection accuracy. Secondly, lightweight depthwise separable convolution (DWConv) replaces standard convolutions in the four downsampling layers of YOLO11’s backbone to reduce computational load and boost detection speed, making it more suitable for UAV real-time detection. Comparative experiments on a self-built dataset show that GWPV-YOLO achieves a $\mathbf{2. 0 \%}$ improvement in $\mathrm{mAP} \text{@} 0.5$, a 26.9% reduction in parameters, a $\mathbf{2 3. 4 \%}$ decrease in FLOPs, and a $\mathbf{2 7. 3 \%}$ reduction in model size compared to the YOLO11 baseline.